MANAGING UNCERTAINTY IN THE DETECTION OF CHANGES FROM
TERRESTRIAL LASER SCANNERS DATA
Reem Zeibak,, Sagi Filin
Dept. of Transportation and Geo-Information Eng., Technion – Israel Institute of Technology, Haifa 32000, Israel
- (reemz, filin)@tx.technion.ac.il
Commission V, WG 3
KEY WORDS: Terrestrial laser scanning, Change detection, Point clouds, Uncertainty
ABSTRACT:
We present in this paper an algorithm for the detection of changes based on terrestrial laser scanning data. Detection of changes has
been a subject for research for many years, seeing applications such as motion tracking, inventory-like comparison and deformation
analysis as only a few examples. One of the more difficult tasks in the detection of changes is performing informed comparison when
the datasets feature cluttered scenes and are acquired from different locations, where problems as occlusion and spatial sampling
resolution become a major concern to overcome. While repeating the same pose parameters may be advisable, such demand cannot
always be met, thus calling for a more general solution that can be efficient and be applied without imposing any additional
constraints. In this paper, we propose a general detection strategy and analyze the actual effect of error sources and artifacts
associated with laser scanning, particularly terrestrial based. The focus of this analysis is not only on the actual sources but on their
interaction when two or more scans are compared. Finding an adequate representation and deriving an error aware model leads to an
efficient and reliable scheme for detecting changes.
1. INTRODUCTION
Terrestrial laser scanning emerges as a mapping technique
capable of providing rapid and direct description of 3D
geometry independent of lighting conditions, and without the
need for a manual collection of the data. The point-cloud
provided by laser scanners is both dense and accurate, thereby
allowing a detailed description of objects irrespective of their
shape complexity. It is therefore, not surprising that laserscanning technology is rapidly becoming the popular alternative
for modeling 3D scenes, for site characterization, cultural
heritage documentation and reverse engineering, as only a few
examples.
The emergence of terrestrial laser scanning as a tool for dense
and accurate 3D mapping has also seen a growing number of
applications in the context of change detection. These usually
refer either to deformation analysis or to quantification of
differences between epochs. For deformation analysis, such
methodologies are usually being implemented by adjustment to
surface models like cylinders (Gosliga et al., 2006) and planes
(Lindenbergh and Pfeifer, 2005) or being based on specifying
the deformation direction in advance (Schäfer et al. 2004). For
the detection of changes between scans on a larger scale, Hsiao
et al. (2004) make use of a data conversion scheme in which 3D
dataset is transformed into a 2D grid. While being a technically
efficient solution, it may lead to loss of valuable information.
Girardeau-Montaut et al. (2005) propose a point based
comparison method that is based on returning to (or close to)
the same scanner position. The need to return to the same
scanner position imposes however, severe limitations on the
detection of changes that more often than not, cannot be applied.
Alterations in the scene, as well as the fact that scanners are not
placed on top of control points (as with theodolites) suggest that
the scanner placement will be in most cases arbitrary or reflect
the present scene configuration. Additionally, a site survey does
not always consider future repeats and so the scan is performed
with scene coverage in mind rather than scanner placement as
close as possible to the previous location.
501
Aiming towards a general change detection scheme, we seek a
methodology that does not impose any external constraints other
than overlap between the scans. The sought model should be
reliable, handle the geometric implications of scanning from
different positions, and be applicable for general cluttered
scenes, as natural scenes tend to be. It should also be
computationally efficient with no imposition of elaborate
processes with added computational overhead. The foundations
of the model presented here have been described in (Zeibak and
Filin, 2007) focusing on the geometrical aspect of the problem.
We study here artifacts and error sources associated with the
scanning process and their effect on the detection. This allows
providing a more thorough understanding about the limit of
detection, and so distinguishing among actual changes and ones
relating to the acquisition system and pose parameters. This
realization, therefore, leads to a more reliable and robust
detection procedure. Among the different inaccuracy sources,
we focus on the effect of objects geometry, objects boundary,
their location and orientation with respect to the scanner, and
their impact on errors in the scan. While similar analyses,
relating to ranging uncertainty have been addressed in the past
(e.g., Lichti et al., 2005), here the association between different
scans must come into effect. We study therefore how errors
originating from individual scans affect the detection of mutual
changes.
The paper is structured as follows: Section 2 presents the
change detection model, Section 3 discusses uncertainties and
error sources associated with laser scanners, analyzes their
effect on the detection of changes and compensate for their error
effect. Section 4 presents results of the application and the
model, and Section 5 offers concluding remarks and outlook.
2. THE PROPOSED MODEL
When studying the detection of changes between terrestrial laser
scans, concerns like data characteristics, level of comparison,
and scene complexity, are key factors that affect the detection
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
strategy. Regarding the level of comparison, it may be applied
at the point level by comparing a point to its surrounding, at the
feature level via primitive based comparison (e.g., planes or
conics), and the object level by comparing objects and their
shape variation between epochs. For an efficient model, we
adopt a point based change detection model, opting towards a
model that does not require elaborate preprocessing stage, e.g.,
segmentation or object extraction that incur high computational
overhead.
2.1 Data representation
3D laser scans can be considered as range panoramas whose
axes are the latitudinal and longitudinal scanning angles, and the
ranges are the intensity values. As the angular spacing is fixed
(defined by system specifications), regularity becomes an
established property of this representation. Relation between the
Cartesian and the polar data representation is given in Equation
(1).
A naive point-based comparison in 3D space will be
implemented by searching for a point’s counterpart within a
spatial neighborhood. Such comparison faces however hurdles
due to the irregular point distribution, variation in scale, and
huge data volume in each scan. The varying resolution within
each scan and between scans inflicts on the level of detail in
which individual objects are described and requires an adaptive
point neighborhood definition. Additionally, parts of the field of
view will oftentimes be occluded by objects (walls most often)
that block their line of sight. While, parts of some regions are
occluded in one scan, they can be visible in the other, and so
might wear a “change-like” form in a point-to-point comparison
scheme. Such features affect the ability to compare the two
scans in a naive neighborhood-search manner, and require
introducing an object notion into the detection.
(x, y, z )T = (ρ cos ϕ cos θ , ρ cos ϕ sin θ , ρ sin ϕ )T
(1)
with x, y and z the Euclidian coordinates of a point, φ and θ are
the latitudinal and longitudinal coordinates of the firing
direction respectively, and ρ is the measured range. Δθ, and Δφ,
the angular spacing, define the pixel size. Figure a shows range
data in this representation where the x axis represents the θ
value, θ ∈ (0,2π], and, the y axis represents the φ value, with φ
∈ (-π/4,π/4] for this scan.
The arrangement of the irregular 3D point cloud into a
panoramic form offers not only a compact representation (which
is less of a concern here) but more importantly an organization
of the ranges according to their viewing direction. To some
degree, this representation can be viewed as tiling of the data,
where the pixel size in angular terms defines a region where the
measured range is the best information source. This contributes
to the connectivity notion as featured in Figure a. Since size is
defined here in angular terms, the varying distance between
consecutive points and scan-lines cease being a factor.
Change detection between scans can also be approached by
asking whether a point that was measured in one scan can be
seen by the other scanner. Three cases can arise from such a
comparison scheme: i) yes, the point can be seen, but there is no
counterpart object in the reference scan, namely a change, ii)
yes, the point can be seen, and it is lying on an object, namely
no change, and iii) no, the point cannot be seen, as there is an
item hiding it, and due to lack of any other information we mark
it as no change. Approaching the detection problem in this
manner requires a different representation of the 3D point cloud.
The following sections present the representation and the
derived detection model.
To assess if a point that was measured in one scan can be seen
by the reference one, the evaluated scan should be transformed
into the same frame as the reference scan. Even though the
range dataset are assumed registered, each range panorama is
given in a local scanner coordinate system.
a.
b.
Figure 1: Range panorama applied for the detection of changes
a. Reference scan. b. Analyzed scan transformed into the reference frame.
502
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
with dij pixel in the difference image, Rij pixel in the reference
image, Aij pixel in the analyzed one, and ε an accuracy threshold.
This image subtraction in the range panorama representation has
some appealing properties: i) when a scan is transformed into
the view point of the reference scan, occluded areas of the
analyzed scan become "no-information" (or void) regions, as
Figure b shows and therefore have no "change-like" effect; ii)
regions seen from the analyzed scan but not from the reference
scan (occluded by the reference) will fall behind the occluding
objects. As such, they have bigger range values than those of
the occluding objects and, therefore, a "no change" status, and
iii) scale – since objects close to the scanner position occupy
more pixels and far objects occupy less, the need to characterize
multi-scale comparison arises. However, as objects from the
analyzed scan are transformed into the reference scan frame,
object scale differences will be resolved in large.
Having the exterior orientation parameters (position and
orientation) of both scans, it is possible to obtain their relative
orientation in order to ease their transformation into a common
reference frame. Given the transformation models for the two
scans
x w = R1x1s + t1
(2)
x w = R 2 x 2s + t 2
with t1 , R1 , t 2 , R 2 the position and orientation matrices of the
two
scanners,
x 2s = ⎡⎣ X s2
Ys2
respectively,
and
x1s = ⎡⎣ X s1
Ys1
Z s1 ⎤⎦
t
,
Z s2 ⎤⎦ the 3D coordinates of a given laser point in
t
As can be noticed, the change detection is directional, namely
negative differences, which are due to occlusion, cannot be
described as changes. Additionally, "no-data" (no-information)
regions in the transformed scan cannot be interpreted as a
change (for consistency with the above definition and for
implementation purposes, we assign "max-range" values there).
So, in order to assess the overall difference between two epochs,
changes should be mutually detected between the different
scans, with the comparison between the reference scan and the
analyzed scan telling what appears in the analyzed but not in the
reference, and the reverse comparison telling what appears in
the analyzed scan but not in the reference. Their union
comprises the overall change; their exclusion reveals the static
objects in the scene.
the individual scanners local frames, as given by Equation (1),
the relative orientation between the two scans can be formed.
This orientation should be formed in a way that the reference
scanner is located at the origin of the common Cartesian frame
and that its orientation is set to be zero, and that the analyzed
scanner will be located and oriented relatively to the reference
one.
Representing a point sampled by the analyzed scanner in the
reference frame will then be given by
x1s = R1t R 2 x 2s + R1t (t 2 − t1 )
N
R 2′
(3)
t 2′
2.3 Point proximity
Despite the simplicity and efficiency of range subtraction, some
adaptations that relate to potential errors in the ranging process
need to be made. These relate to some sensitivity in cases where
the incidence angle between the scanner and the surface
becomes high.
and so, transforming a point from the analyzed scan to the
reference frame has the form of a rigid body transformation
with R 2 ' the relative rotation and t 2 ' the relative translation.
The relative transformation of the analyzed scan into the
reference scan coordinate system allows working in the same
coordinate frame, a property that disappeared with the
formation of the range panoramas. A subsequent step of this
transformation will be computing the pointing angle and range
N
2
Δh
This transformation leads to an angular correspondence between
the reference and the transformed analyzed scans.
Applying this transformation to the whole analyzed scan will
then have the notion of asking how the scan (namely the data
acquired in the scan) looks from the reference scanner position.
Figure b shows the analyzed scan as transformed into the
viewing point of the reference scanner in Figure a.
In points where the incidence angle α between the beam and
the surface normal is relatively big, evaluating point proximity
between counterparts becomes sensitive as the beam is almost
parallel to the object surface and therefore leads to greater
uncertainty in the ranging Δρ (see Figure 2). Variation in the
scanning direction and thus in the incidence angles, would result
in spatially different points that describe the same object.
However, the distance between the corresponding points along
the surface normal direction in the reference frame Δh provides
a more subtle evaluation for point proximity in the detection of
changes. Therefore, some modifications are made for points
resulting in big difference of incidence angles in both scans, by
testing the distance between the observed counterparts in the
direction of the surface normal, and deciding whether the point
When transformed, comparison between the scans can be
reduced with some adaptations, into a mere image subtraction.
The detection of changes will then be based on the following:
di, j
Ri , j − Ai , j > ε
Ri , j − Ai , j ≤ ε
Ri , j − Ai , j < −ε
α
Figure 2: Range difference in reference frame, projected to the
surface normal direction.
2.2 Detection of changes
⎧ change
⎪
= ⎨ no change
⎪occlusion (no change)
⎩
Δρ
Reference
to the measured point x s as seen from the reference scanner.
(4)
503
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
is far from the pre scanned surface or not. This modification
requires an efficient normal vector computation that needs to be
performed only on the reference scan, as points in the analyzed
scan are compared to what have been scanned in the reference
and not vice versa.
Pole
3. LASER BEAM UNCERTAINTY AND ERROR
SOURCES
Figure 4: Demonstration of artifacts due to sliver points along
the tangency plane of an object.
So far, a geometrical interpretation of the detection of changes
has been proposed. Nonetheless, errors associated with the
scanning process exist and unless treated, affect the quality of
the results by increasing the level of false alarms. Errors can
largely be divided into random components associated with
measuring process, and artifacts.
Figure 4 provides one example demonstrating those artifacts,
showing the point cloud of a pole scanned from close range
(about 2-3m). Other than the pole's front, phantom points can be
seen around it, relating to transmitted pulses that are tangent to
its borders. Other than proximity to boundaries, the geometrical
structure of an object can cause similar artifacts. For example,
holes within bodies scanned close to the scanner (about 2-3m)
result in outliers in the scanning direction (see figure 5).
3.1 Random error sources
Random errors in raw laser measurements lead to uncertainty in
3D point position and therefore need to be modeled. The
significant error components in the laser scanning systems refer
to the ranging σ ρ [m] and to the recording of angular firing
direction σ ϕ , σ θ [rad ] . A proper error propagation of such
Chain
uncertainties can provide a theoretical estimate for the threshold
value of the range difference, ε , needed in the proposed change
detection.
3.2
Artifacts
The random error budget features only one aspect of the laser
related inaccuracies. Other elements are termed here artifacts
since they are manifested as outlying points within a point cloud
but exhibit neither random nor systematic trend. Their causes
stem from a variety of sources, and their magnitude varies from
one case to the other. Discussion about artifacts that
characterize laser scanners can be found in Lichti et al. (2005),
Sotoodeh (2006), and Barnea and Filin (2008). These data
artifacts wear different effects like: angular displacement, mixed
pixels, saturation, blooming, multi-path and no-reflectance
points.
Figure 5: A shower of outlying points formed due to a hollow
structure and a chain form.
In terms of their effect on the change detection process, a point
based model that compares points in one scan to their
counterparts in the other, can assign those points a “changelike” tag. This realization has to do with the fact that they do not
refer to the observed objects structure in reality (see figure 6)
and therefore cannot be scanned twice. Consequently they will
not find counterparts in other scans.
Chain
Pole
Figure 6: Results showing the data following the elimination of
edge related artifacts.
Figure 3: Sliver points formed along the tangent planes to a
scanned object.
Most error sources tend to appear around tangency lines or
planes to the surveyed objects. Due to the laser beam width,
ranging artifacts mainly appear in discontinuity regions; where
from a single ranging, multiple returns of different objects occur.
The ranging artifacts linked to object contours (termed also
edge effect) lead to outlying points that appear in the data but do
not exist in reality and cannot be referred to any object structure
(See figure 3).
504
Figure 7 illustrates this scenario where the cylinder shape
element in front of the "red" scanner causes phantom points to
follow its tangent plane (as in figure 4), while a scanner placed
in the "blue" position will only view the wall in front of it.
Comparing "red" scan to the "blue" one shows the entire
phantom points as actual changes, as they are placed in front of
the blue object (the wall) and have no counterpart points there.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
The analysis begins with testing the raw range based
comparison process. This phase involves only the
transformation of the analyzed range panorama into the
reference one and performing image subtraction as presented in
Section 2.2, the only parameter that influences the results is the
value of ε.
Reference
“Change-like”
points
a.
Figure 7: Change like effect associated with phantom points.
When transforming the analyzed scan into the reference frame,
outliers nearby edges in the analyzed scan will be transformed
as well. And thereby will be compared to inliers in the reference.
To differentiate between actual changes and those outliers,
compensation for the edge effect is applied.
In order to compensate for such edge effect, a minimum range
difference is searched in the point’s neighborhood during the
range subtraction. This way, it is compared to other points
belonging to that object in the reference scan and not to other
objects falling way behind. The proposed solution for reduction
this kind of edge effect, involves pre detection of edges in the
analyzed scan using Laplacian of a Gaussian (LoG) filter.
b.
Figure 9: a) results of applying the change detection with
ε=5cm for range difference (window 1 in figure 8), b)
results with Δh=5cm for normal distance.
Figure 9a shows that applying a threshold value of 5cm for
range difference yields a large amount of ground returns that are
categorized as changes in those regions due to relatively big
incidence angles there in the reference scan. Therefore, a
change detection scheme that is based on a mere range
subtraction for such cluttered scenarios without filtering the data
exhibits sensitivity to ranging uncertainty as a function of
incidence angle (see Section 2.3). To reduce the false changes
resulting from the naive range-difference based comparison,
analysis of changes along the normal direction is applied as
proposed in Section 2.3. The threshold value Δh is set to 5cm.
This mode of comparison is applied to all range differences
smaller than 50 cm and whose incidence angle is large. The
results of this modification are shown in Figure 9b. The
improvement in terms of amount of false alarm detections is
substantial. Other than only demonstrating the effect of the
proposed modification, it tells about laser ranging process. We
note that in reference to the results in Zeibak and Filin (2007),
the use of a dilation-like operator in the form of locally dilated
close objects and eroding the background enabled eliminating
such errors. However, this process was too global when applied
to the whole scan.
4. RESULTS AND DISSCUSION
The application of the change detection model is demonstrated
in Figure on a complete scene (the analyzed scan is compared
here with the reference). Changes in this scene between the two
epochs are mainly passersby and parking vehicles. They
indicate the level of detail that can be noticed at various ranges.
Considering the very different views from which the two scenes
were acquired (about 11 meters distance between scan
positions), the ability to detect walking persons shows the great
potential of the proposed approach to detect changes of various
size and within a cluttered natural environment while managing
occlusion and laser scanning related artifacts effects between
the scans.
Since the proposed model involves different phases, beginning
with the overall visibility query concept and then adapts itself to
the ranging related features and artifacts, we evaluate the
contribution that each of these processes has on the comparison.
To illustrate the different phases, we focus on three different
regions in the scan that best demonstrate the contribution of
each phase (see Figure 8).
Figure 8: Changes detected in the analyzed scan when compared to the reference one.
505
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
Next, figure 10a and 10b show the results using the range
subtraction and normal distance test (Δh =5cm) in a region
featuring a significant amount of discontinuities. One can notice
the very many false alarms appearing on static objects, close to
their edges, e.g., windows boundaries, poles, signs, etc. This
edge effect increases further in regions with poor registration
accuracy.
Detailed study of the error sources and their effect on the
detection of changes shows that unless considered, a large
amount of false alarms may result. However, when accounted
for, a relatively small detection threshold (like the 5 cm value
set here) can be applied.
REFERENCES
a.
Girardeau-Montaut, D., Roux, M., Marc, R., Thibault, G., 2005.
Change detection on points cloud data acquired with a ground
laser scanner. International Archives of Photogrammetry,
Remote Sensing. 36(3/W19):30-35
b.
Gosliga van R., Lindenbergh, R., Pfeifer, N., 2006. Deformation
Analysis of a Bored Tunnel by Means of Terrestrial
Laserscanning. Image Engineering and Vision Metrology,
ISPRS Symposium, Dresden.
c.
d.
Hsiao, K. H., Yu, J., Tseng, Y. H., 2004. Change Detection of
Landslide Terrains Using Ground-Based Lidar Data.
International Archives of Photogrammetry and Remote Sensing,
35(B5)
Lindenbergh, R., Pfeifer, N., 2005. A Statistical Deformation
Analysis of Two Epochs of Terrestrial Laser Data of a Lock. 7th
Conference on Optical 3-D measurement techniques, Vienna.
Figure 10: a, b) Results of applying the change detection with
ε=50cm, ε=5cm (windows 2,3 in Figure ). c, d) applying the
change detection with the elimination of the edge effects.
Schäfer, T., T. W., Peter Kyrinovič, Miriam Zámečniková, 2004.
Deformation Measurement Using Terrestrial Laser Scanning at
the Hydropower Station of Gabčíkovo. INGEO 2004 and FIG
Regional Central and Eastern European Conference on
Engineering Surveying. Bratislava, Slovakia
Following the application of the edge effect correction, spurious
detected changes relating to edge effect are eliminated. In
Figure 10c and 10d the results of eliminating edge related
artifacts can be seen. Notable false alarms that have been
removed can be seen with many of the poles and particularly on
edges. As with the modification to the range subtraction, the
ability to model the effect of along-edge points is exhibited by
their elimination.
Zeibak, R., Filin, S., 2007. Change detection via terrestrial laser
scanning. International Archives of Photogrammetry and
Remote Sensing. 36(3/W52): 430-435.
Lichti, D.D., Gordon, S.J., Tipdecho, T., 2005. Error Models
and Propagation in Directly Georeferenced Terrestrial Laser
Scanner Networks. Journal of Surveying Engineering, Vol. 131,
No. 4, November 1, 2005.
5. CONCLUSIONS
Sotoodeh, S., 2006. Outlier detection in laser scanner point
clouds. International Society of Photogrammetry and Remote
Sensing.
This work has demonstrated the feasibility of change detection
with no imposition of external constraints. The results
emphasize the capability of the proposed model to detect
changes in general and cluttered scenes via terrestrial laser
scanning irrespective of the scanning positions, showing its
great potential. It has also shown that multi-scale objects
appearing in different depths are successfully detected. Also
dynamic objects that pass through the scene during a single scan
could be definitely alarmed as changes.
Barnea, S., Filin, S., 2008. Registration of Terrestrial Laser
Scans via Image Based Features. ISPRS Workshop on Laser
Scanning 2007 and SilviLaser 2007, Espoo, September 12-14,
2007, Finland.
506